In Defense of (Human-led) Science
In Defense of (Human-led) Science
Chirag Shah, University of Washington
Last year, researchers at Sakana AI unveiled something remarkable: an AI system that could generate a novel research idea, design experiments to test it, run those experiments, analyze the results, write up the findings as a complete scientific paper, and even review its own work — all for about fifteen dollars. Just a few months ago, Google announced its AI co-scientist, which reportedly compressed what took human researchers a decade of iterative investigation into roughly two days of computation.
The message seems clear: science can now be manufactured at scale.
And to some, this is exactly the point. Why maintain expensive labs full of graduate students and postdocs when algorithms can churn through hypothesis space with tireless efficiency? Why wait years for breakthroughs when AI can accelerate the clock speed of discovery itself? The economic logic appears unassailable. The technological capability is demonstrably real.
But something essential gets lost in this framing — something that the optimists of automated science have either overlooked or chosen to ignore.
The Confusion of Product and Process
Here is the fundamental error: treating science as if it were merely about its outputs. Papers published. Hypotheses confirmed. Drugs discovered. When we talk about “accelerating science,” we almost always mean accelerating the production of these artifacts. But doing science is not the same thing as producing science.
Science, at its core, is a profoundly human activity. It involves not just generating knowledge but developing judgment about what questions matter. It requires navigating ethical terrain that cannot be reduced to optimization functions — asking not only how something might be done, but whether it should be done at all. It demands dialogue: between researchers and their critics, between scientists and the communities affected by their work, between disciplines that must learn to speak across their boundaries.
—When we talk about "accelerating science," we almost always mean accelerating the production of these artifacts. But doing science is not the same thing as producing science—
An AI system can simulate the structure of peer review. It cannot replicate the years of accumulated wisdom that allow a seasoned reviewer to recognize a promising but flawed approach, or to encourage a struggling graduate student while redirecting their efforts. These capacities emerge from sustained practice, from making mistakes and reflecting on them, from participating in a community of inquiry over time.
The Deskilling Trap
There is a growing body of evidence that when we automate cognitive tasks, we do not merely offload work — we erode the very capacities that made us effective in the first place. Pilots who rely too heavily on autopilot lose their ability to handle emergencies. Radiologists who depend on AI detection systems become worse at spotting anomalies when the system fails. Medical students trained with unrestricted AI assistance underperform their peers once that assistance is removed.
The pattern is consistent: the more a tool does for us, the less capable we become without it. This is not merely about practical competence; it extends to what researchers call tacit knowledge — the deeply embedded, often unarticulated expertise acquired through hands-on practice and iterative failure. AI systems, by automating the very activities through which such knowledge develops, remove the experiential foundation on which expertise is built.
Applied to science itself, the implications are sobering. If we train a generation of researchers who never struggle through the manual labor of literature review, who never experience the frustration of a failed experiment that nonetheless teaches them something crucial about their system, who never sit in a room arguing about the interpretation of ambiguous data — what kind of scientists will they become? What will they be capable of when the AI fails, or when they face a genuinely novel problem that lies outside the training distribution of any existing model?
What We Stand to Lose
The risk is not that AI will replace scientists. The risk is that we will hollow out the scientific enterprise from within, preserving its outward form while emptying it of substance.
We will still have papers — more than ever, perhaps. We will still have hypotheses and discoveries and Nobel Prizes. But we may lose the capacity to critically evaluate what those papers actually mean. We may lose the judgment to know when a line of research should not be pursued, regardless of what the optimization function recommends. We may lose the ability to take genuine responsibility for scientific work, because responsibility requires understanding, and understanding requires engagement.
—The risk is not that AI will replace scientists. The risk is that we will hollow out the scientific enterprise from within, preserving its outward form while emptying it of substance—
Science has always been more than a knowledge-production machine. It is a practice that transforms those who participate in it. The graduate student who emerges from a PhD program is not simply someone who has produced a dissertation; they are someone who has been shaped by the discipline of inquiry, who has learned to think differently, who has developed intellectual virtues that cannot be acquired any other way. This transformation is not a side effect of doing science — it is part of what makes science valuable.
When we reduce science to its outputs, we treat this transformation as waste. We see the years spent developing expertise as inefficiency to be optimized away. But those years are not merely instrumental; they are constitutive of what it means to be a scientist.
The Case for Human-led Science
None of this is an argument against using AI in scientific research. The tools are genuinely powerful, and in the hands of skilled researchers, they can accelerate certain aspects of inquiry while freeing human attention for questions that require it most. A scientist who uses AI to rapidly survey a literature, generate candidate hypotheses, or analyze massive datasets may be more effective precisely because they retain the judgment to know what to do with those outputs.
The critical distinction is between AI as instrument and AI as replacement. An instrument extends human capability while leaving the human in control. A replacement removes the human from the loop entirely.
What we need is not automated science but augmented scientists — researchers who remain firmly in the lead, who use AI tools strategically without ceding the essential activities through which scientific expertise and judgment develop. We need to resist the seductive efficiency of full automation and insist that the messy, slow, sometimes frustrating process of human scientific inquiry is not a bug to be fixed but a feature to be preserved.
This will require deliberate choice. The economic pressures pointing toward automation are real. The technological capabilities enabling it will only increase. Maintaining human-led science in this environment will demand institutional commitment, funding structures that value process alongside output, and a willingness to accept that some things worth doing cannot be optimized.
But the alternative — a world in which science is produced but no longer practiced, in which we have papers but not understanding, discoveries but not discoverers — is not merely less efficient. It is impoverished in ways that matter more than any efficiency gain could compensate.
Science is a human activity. Let us keep it that way.
Cite this article in APA as: Shah, C. (2026, February 11). In defense of (human-led) science. Information Matters. https://informationmatters.org/2026/02/in-defense-of-human-led-science/
